Behavioural Model for Community-Based Antimicrobial Resistance, Vellore, India

1. Behavioural Model for Community-Based Antimicrobial Resistance, Vellore, India Dele Abegunde1; Holloway, Kathleen1; Mathai, Elizabeth1; Gray, Andy2; Ondari, Clive1; Chandry, Sujith3 1World Health Organization, Switzerland; 2Nelson Mandela School of Medicine, University of KwaZulu-Natal, South Africa; 3Christian Medical College Hospital, Vellore, India

2. Background Resistance to antimicrobial agents compounds the burden of diseases worldwide. Difficulties to estimating the impact of AMR on individuals and the community or the impact of AM use on resistance in resource-constrained settings is compounded by the paucity of community-based data. Robust surveillance data collection methodologies are lacking in such settings. More explorations and improved analytical methods are needed to fully understand trends and impact of AMR on cost of illness and to inform AMR surveillance.

3. Empirical USE-Resistance-USE model USE Resistance Health Impact Economic Impact Appropriate & inappropriate use of Antimicrobial agents Antimicrobial Resistance Prolonged morbidity Cost of treatment cost of laboratory investigations Increased risk of mortality risk of mortality Increased risk of complications loss productivity & visit costs Health system cost Transference Ecological model Until this exploration, we have found no analysis in the literature that adequately or directly accounts for reverse causality or endogeniety

4. Objectives • To determine the behavioural trends—seasonality (periodicity) and the temporal associations—between community-based AMR and AM use; • To forecast the short-run pattern in AMR through the behaviour of AMR and the predictors (indicators) of AMR; and • To compare the temporal correlation of the trends in DDD and the proportion of patients prescribed antibiotics, with community based AMR • Explore improved methodological perspectives in analyzing ecological AMR

5. Methods Data • AMR surveillance data from Vellore (urban area) and KV Kuppam, situated between Chennai and Bangalore • combined population of 500, 000 in a 3.5million Vellore district in the state of Tamil Nadu, Southern India. • AMR surveillance data consist of commensal E. coli isolated from urine/perinea (swab) samples obtained from asymptomatic pregnant women attending antenatal clinics. • Monthly AM-use data were those obtained from exit interviews conducted by pharmacists from urban and rural facilities: • hospitals or primary care clinics (including not-for-profit and for-profit hospitals in the rural and urban areas); • private sector pharmacies; and • private sector general medical practitioners’ practices. • All data were collected in two-time period, from August 2003 to July 2004 and from January to December 2005.

6. Methods Variables • AMR data was converted into proportion of the total E coli isolates, which were resistant to a specific antibiotics class: • co-trimoxazole, • extended spectrum penicillin (ESP) and • quinolones (nalidixic acid and fluoroquinolones). • The AM-use data were: • standardized to DDD of the respective antimicrobials and • the proportion of prescriptions containing specific antimicrobial groups within the total prescriptions for the month.

7. Methods: Analysis Models: Autoregressive Integrated Moving Average (ARIMA) Univariate Assumes causal links Xt = M x 1 Vector of exogenous variables – antimicrobial use in monthly total of DDD or Proportion receiving antibiotics, and β is a K x M matrix of coefficients, Ρ = first order autocorrelation parameter θ = first order moving average parameter Ɛt = white noise ~ i.i.d. N(0, δ2) Vector Autoregressive Analysis (VAR) Multivariate Allows for examination of causality Yt= Proportion of AMR in month t, (Y1t …… Ykt) is a K x 1 random vector of lags and the ƞi are fixed K x K matrices of parameters, δ = Constant – K x 1 vector of fixed parameters, µt = The disturbance term assumed to be the white noise, P = lags of Yt, and t = month. Holts-Winters seasonal smoothening technique for trends

8. Figure 2: Trends in antibiotic use variables compared.

9. Actual and Predicted trend in resistance

10. Predicted trend in community-base antimicrobial resistance and antimicrobial use

11. Impulse response function

12. Results Table 1 : Granger causality tests

13. Summary Both AMR and AM-use demonstrated lagged trends and seasonality. Parameter estimates from the VAR (table 2) are more efficient compared to those form ARIMA (table 1). Seasonality spurs of resistance appear to synchronise with cold (catarrh) seasons when the antibiotics are freely and routinely used. AMR lags vary between 3-5 months of AM-use. This also synchronizes with the cold periods. AMR trend is sustained even though antibiotic use trends downward. Impulse-response could last as much as 15 to 45 months. Indicating that AMR resistance generated by a bout of inappropriate use can last in the communities for up to 15 -45 months. AM-use demonstrated significant Granger causality with AMR in addition to circularity. Both monthly DDD per patient and proportion of patients on specific antibiotics show similar effects on AMR, but DDD per patient appear to demonstrate more reactive effect on AMR.

14. Summary of findings • Refined models provide clearer knowledge of the dynamic and systematic relationships between antibiotic use and antimicrobial resistance in respective communities. • Community AM use can predict AMR. • Linearized models are scientifically and empirically intuitive, and are useful tools for forecasting, monitoring and evaluating future deviating observations • Estimating parameters to support robust policy and survielance designs requires the use of more robust analytical methodologies. • Results provide additional evidence for estimating the economic impact of AMR and could inform the design of community-based antimicrobial surveillance and interventions in low-resource settings. • Results provide evidence to support the of utility of cheaper-to-measure antibiotic-use variable

15. Table 1: Autoregressive Integrated Moving Average Regression (ARIMA) results

16. Table 2: Vector Autoregression (VAR) results

17. Table 2: continues

18. Thank you for your patience in listening